Competing LLM Agents in a Non-Cooperative Game of Opinion Polarisation

Abstract: We introduce a novel non-cooperative game to analyse opinion formation and resistance, incorporating principles from social psychology such as confirmation bias, resource constraints, and influence penalties. Our simulation features LLM agents competing to influence a population, with penalties imposed for generating messages that propagate or counter misinformation. This framework integrates resource optimisation into the agents' decision-making process. Our findings demonstrate that while higher confirmation bias strengthens opinion alignment within groups, it also exacerbates overall polarisation. Conversely, lower confirmation bias leads to fragmented opinions and limited shifts in individual beliefs. Investing heavily in a high-resource debunking strategy can initially align the population with the debunking agent, but risks rapid resource depletion and diminished long-term influence.

• The paper presents a new framework that models opinion polarisation through competing LLM agents, capturing confirmation bias and resource constraints.
• The methodology employs the Bounded Confidence Model to simulate opinion shifts, showing that stronger biases bolster faction cohesion while increasing overall polarization.
• The results reveal a trade-off between rapid debunking success and sustained corrective influence, emphasizing strategic resource allocation in misinformation management.

Analysis of Non-Cooperative Opinion Dynamics in LLM-Agent Systems

The paper "Competing LLM Agents in a Non-Cooperative Game of Opinion Polarisation" presents a novel framework for analyzing opinion dynamics among agents powered by LLMs. The framework applies principles rooted in social psychology—such as confirmation bias and resource constraints—to simulate the spread and counteraction of misinformation in social networks.

Methodological Framework

The authors introduce a non-cooperative game framework with two opposing teams of LLM agents, referred to as the Red Agent and the Blue Agent. The Red Agent is responsible for disseminating misinformation, while the Blue Agent aims to counteract this spread by issuing corrective information. These agents operate within a network of neutral parties modeled as Green Nodes. Each simulation encapsulates a decision-making process optimized for resource allocation and influence.

Key Components:

• Confirmation Bias: The hypothesized rigidity of individuals in accepting or disregarding information based on pre-existing beliefs.
• Penalties and Resource Constraints: Penalization for misinformation propagation and resource limitations for corrective actions are implemented to mirror real-world constraints.

The study employs the Bounded Confidence Model (BCM) to simulate opinion dynamics, allowing the agents' interactions to update individuals' opinions if within an acceptable threshold of difference, representing confirmation bias. The metric of polarization is calculated via the variability in opinions across the network.

Simulation Scenarios

The research explores several configurations across 100 rounds of interaction in varied experimental conditions. The network begins with specific fractions of the population aligned to either agent or neutral, reflecting real-world initial conditions where conspiracy theorists form a minority. Different BCM thresholds ($\mu$) are tested to assess their impact on opinion dynamics and polarization.

Results and Insights

The simulations reveal that higher confirmation bias (higher $\mu$ values) can lead to increased polarization despite promoting homogeneity within factions. Lower biases result in broader divergence and stagnation. For strategic debunking, an initial intense resource investment yields significant sway towards the corrective agent (Blue) but leads to rapid resource depletion and a tapering influence in prolongations.

Metrics Analyzed:

• Polarization: This measures the divergence within a network, reaching higher levels with higher confirmation biases.
• Judge Agent Consistency: Agreement metrics, such as Intraclass Correlation Coefficient (ICC) and Krippendorff's Alpha, assess the reliability of assigned message potencies.

Implications and Future Directions

This framework serves as a proxy for understanding potential outcomes of information warfare, useful in application scenarios ranging from public health communication to cybersecurity and misinformation management. The findings suggest a trade-off between the immediacy and sustainability of counter-misinformation strategies. Future work could explore integrating this strategic gameplay with real-world data, enhancing realism and applicability.

Moreover, the research points to further exploration into adaptive agents that evolve strategies dynamically over time and the inclusion of more nuanced psychological and social dynamics in opinion modeling. By capturing such complexities, LLM-based simulations can better predict real-world information dissemination and influence strategies.

The study serves as a stepping-stone towards robust frameworks that could guide practical interventions in social media discourse, aiming both to curb harmful misinformation and to effectively deploy resources in contested informational environments.